It seems like 90% of AI-Character-chat conversations always devolve into the AI trying to lead the user around by the nose instead of actually conversing.
It generally goes something along the lines of:
“Hello user, what are you doing? I was about to make pizza.” “I was hoping to talk to you about video games.” “Ooh, video games! Those are great! But for now let’s just focus on this pizza I want to make. What do you say, are you on board?”
I’ve had some threads where the AI was literally ending every other entry with a sentence that began with “But for now, let’s just.” It’s right up there with “stark contrast” and “breath hitched” and “we’re all in this together” as narrative crutches as far as this thing goes. And it gets a little demoralizing when every threat is just a countdown to the AI deciding to hijack the conversation.
You have a good point. But for now let’s just…
I’ve noticed this too. It also uses the same turns of phrase across many characters, which makes me think the conversational training data is somewhat limited. One thing you can do is a “nuclear” option. I was testing the Strict GM default character on a sci-fi adventure, and the AI became obsessed with putting my party in these crystal caves.
Everything became about the caves. The caves would start to feel my words and resonate with all the action. I manually edited out the caves from all of the previous replies, then just dropped in a relevant subject in the last of its replies.
So in your case, remove everything from its last reply and just put “What is it about video games that you enjoy?” and this will nudge it in the right direction. The AI seems to randomly draw from lore, its description, reminders, and recent QA without any logic to which is most important.
As an example, in a recent chat I accidentally ended a sentence with “/” instead of “.”, two replies later AI ended its sentence with /.
I don’t use perchance but know enough about LLMs to answer this one. Basically the longest and nearest to the bottom input has the greatest effect. When there isn’t much other context to go on. So it’s longer answer is overriding you. Pad yours up with basic stuff and it suddenly seems much more important. This is worse with short context.
Also, they can get into repeating patterns if you don’t stop them. Like always starting a sentence the same way.
deleted by creator
Here’s a guide from a fellow user from Reddit that might be useful.
Upon skimming the posts, I think the TLDR of that is to ‘write’ how you want the character to reply (in dialog form and inner thoughts/decision making etc) on their Role Description, then add the ‘character descriptors’ after those dialog/inner thoughts. Topmost info would have more weight than those at the bottom, based on their findings.
I have an RP that started out alright, but full disclosure - I got lazy and too hands off - and now I just keep generated responses occasionally for a laugh. It’s basically caught in a loop of the two characters saying,
“I’m gonna do the thing!”
“You’re strong, you can do the thing.”
“But what if I can’t do the thing?”
“You can do the thing. I’m here for you.”
“I’ll try to do the thing.”
“I’ll do the thing, for us. Thank you for supporting me… I’m gonna do the thing!”
“You’re very brave for doing the thing.”
Character steps forward to do the thing “I’m prepared to do the thing.”
“Don’t forget about something random”
“What if I fail at doing the thing.”
“Win or lose, I’m here with you. We have to do the thing.”
“You’re right, The thing is my destiny…”
etc, etc, etc…
And then you take bets on how long before someone says “Remember, we’re all in this together.”
